humblebanana / dsh-record-replay

dsh-record-replay

A DeepSeek Harness plugin that turns the Open Record/Replay macOS workflow recorder into first-class harness capabilities.

It registers the open-record-replay skill and six model-facing orr_* tools, so an agent can learn a user-demonstrated desktop workflow: record the user's real macOS actions, validate the evidence, and package it for the host agent's native skill creator.

user demonstrates a workflow
  -> orr_record_start            (capture session.json + events.jsonl)
  -> orr_record_stop             (finalize)
  -> orr_session_validate        (check the evidence against the contract)
  -> orr_session_events          (read what the user actually did)
  -> orr_skill_prepare           (package a skill-input directory)
  -> host skill creator

Requirements

• macOS (the recorder's native backend is Swift; it needs Xcode Command Line Tools).
• Node.js >= 22.19 (the Harness runtime).
• A DeepSeek Harness installation.
• An open-record-replay checkout whose bin/orr.js the plugin invokes.

Install

Build the package and add it to a profile as a bundle:

git clone https://github.com/<you>/dsh-record-replay.git
cd dsh-record-replay
pnpm install
pnpm build
pnpm pack                      # produces dsh-record-replay-0.1.0.tgz
dsh plugin --profile web add ./dsh-record-replay-0.1.0.tgz

dsh plugin add records the package in the profile's package.json dependencies and dsh.profile.bundles, and the harness heals the profiles/node_modules fallback so the bundle resolves. The shipped cordis.patch.yml mounts a neutral row; point it at your checkout by overlaying the row from your profile's cordis.patch.yml:

- id: record-replay
  config:
    repoRoot: '/absolute/path/to/open-record-replay'
    runsOut: 'runs'
    skillInputsOut: 'skill-inputs'

The profile patch file is hot-reloaded, so the running GUI picks the plugin up without a restart. Restart the Harness if you are not on a live profile.

Configuration

Key	Default	Meaning
cliPath	env ORR_CLI_PATH	Explicit path to bin/orr.js. Overrides repoRoot.
repoRoot	env ORR_REPO_ROOT	Path to an open-record-replay checkout; the CLI is <repoRoot>/bin/orr.js.
runsOut	runs	Workspace-relative recordings directory.
skillInputsOut	skill-inputs	Workspace-relative skill-input packages directory.

The CLI runs with the session workspace as its working directory, so recordings and skill packages land where the agent's filesystem tools can read them.

Tools

Tool	CLI mapping	Purpose
orr_permissions_check	permissions check	Verify Accessibility / Input Monitoring before recording.
orr_record_start	record start	Begin capturing the user's demonstration.
orr_record_stop	record stop	Finalize the session after the user finishes.
orr_session_events	session events	Read the evidence stream (events.jsonl), capped at limit events.
orr_session_validate	session validate-recording	Check the recording against the official contract.
orr_skill_prepare	skill prepare	Build the evidence package for the host skill creator.
orr_skill_create	— (built-in)	Built-in skill creation fallback. Generates a SKILL.md skeleton from a recorded session following the Anthropic skills spec (kebab-case name + description frontmatter, progressive-disclosure body, evals/evals.json), validates agent-authored drafts, and installs to ~/.agents/skills/<name>/. Prefer a host-native Skill Creator when one exists; this is the fallback.

Development

pnpm install
node scripts/link-dsh.mjs    # links @deepseek-ai/* from a DSH checkout/harness
pnpm build                   # tsc
pnpm test                    # vitest
pnpm validate                # build + test

Tests assert the registration contract and exercise the CLI runner through a stubbed subprocess service; node scripts/link-dsh.mjs --path <dir> points the @deepseek-ai scope at an explicit DSH checkout.

License

MIT — see LICENSE.